The water math on AI is the most misunderstood statistic in tech, so let's actually run it.
A ChatGPT query uses about 0.3 mL of water for cooling, per OpenAI's own disclosure. The harshest academic estimate, which also counts the water consumed by the power plants feeding the servers, lands at 10 to 25 mL. Call it half a tablespoon at the absolute worst case.
Now the comparisons. One hamburger takes roughly 2,500 liters of water to produce. That single burger equals somewhere between 100,000 and 8 million ChatGPT queries depending on whose estimate you trust. A pair of jeans cost 7,500 liters before it reached your closet. One toilet flush is 6 liters, or about 20,000 queries. US golf courses pour out around 1.5 billion gallons every day, dwarfing the direct cooling water of every data center in the country.
So why does the robot demanding a thousand glasses of water feel true? Because the totals sound enormous and nobody does the division. "Data centers consumed billions of gallons" is accurate and terrifying right up until you divide billions of gallons by trillions of queries and get a number smaller than the drop left in your glass when you think it's empty.
The legitimate water story is local. A big facility sited in a drought-stressed basin can strain that specific community, which is why Chile blocked one. That's a siting problem, solvable with policy and closed-loop cooling.
The cartoon will outlive the correction anyway. Per-unit math never beats a thirsty robot at the foot of the bed.
New Yorker recently. This idea will never die.